Sep 12 – 17, 2021
University of Birmingham
Europe/London timezone

Automatic detection of scintillation light splashes using conventional and deep learning methods

Sep 15, 2021, 1:02 PM
1m
Teaching and Learning Building (University of Birmingham)

Teaching and Learning Building

University of Birmingham

Edgbaston Campus University of Birmingham B15 2TT UK

Speaker

Ms Yangfan Jiang (Loughborough University)

Description

The Hybrid Gamma Camera (HGC) [Lees et al, Sensors, 17(3):554, 2017] has applications in intraoperative imaging guidance and nuclear decommissioning. The HGC gamma detector comprises an Electron Multiplying Charge-Coupled Device (EMCCD) coupled to a columnar CsI: Tl scintillator. Each absorbed gamma photon will produce a scintillation light splash on the EMCCD; the number, location, and size of these splashes on each frame is unknown. For real-time imaging, these splashes must be identified and characterised quickly (10fps) and with minimal processing overhead. The current technique used in the HGC is automatic scale selection [Hall et al, NIMA,604:207-210, 2009]. However, there alternative solutions including object recognition deep learning models may have superior performance.
Six automatic detection implementations - four traditional blob detection algorithms (two Laplacian of Gaussian (LoG) detectors: LoG(1) is our implementation of LoG algorithm based on the HGC’s automatic scale selection and LoG(2) is the LoG detector from Python scikit-image library, one Difference of Gaussian (DoG) detector, and one Hessian of Determinant (DoH) detector) and two Region-Based Convolutional Neural Networks (RCNNs) based on the VGG16 and ResNet-101 models were compared for the task of scintillation light splash identification.
An image dataset was simulated based on statistics of experimental data from the HGC for a range of radioisotopes. This provided ground truth values for location, intensity, and size of expected splashes for training RCNN models. The testing performance of each technique was compared based on results from 360 simulated images. The F1-scores obtained were: LoG(1): 69.98%, LoG(2): 98.85%, DoG: 94.03%, DoH: 32.05%, VGG16: 90.37% and ResNet-101: 89.95% . Their average speed for per frame was LoG(1): 0.358s/f, LoG(2): 0.024s/f, DoG:0.021s/f, DoH: 0.069s/f, VGG16: 0.057s/f, ResNet-101: 0.072s/f. We will discuss the physical parameters that affect algorithm accuracy and performance.

Nationality Chinese
Institute University of Loughborough
Title Ms
Your name Yangfan Jiang
email Y.Jiang@lboro.ac.uk

Primary authors

Ms Yangfan Jiang (Loughborough University) Dr Sarah Bugby (Loughborough University) Dr Georgina Cosma (Loughborough University)

Presentation materials